Index 2026.10b

How the index is scored

Last recomputed Oct 16, 2026. Every input lives in the open — the same data files power the rankings, the profiles and the charts. This page explains the pillars, the grades, how every blank is filled with a labelled estimate, and the factors behind the hardware and materials figures.

Reading a score

What the number, the band and the grade mean

Leader65 and above

Discloses real numbers and runs mostly on clean power. Still has a footprint, but you can see it.

Middle of the pack45 – 64

Some disclosure or clean-energy commitments, offset by rising emissions or blind spots.

LaggardBelow 45

Little or no disclosure, fossil-backed capacity, or active disputes with regulators and communities.

A score is the weighted sum of six pillar scores. Each pillar also carries a confidence grade from A (the company published it, with method) to D (little data), and the company grade is the weight-averaged result. A high score with a D grade means the company looks fine on the little it publishes.

Five kinds of number

Company figure

Published by the company itself. We take it as given and record the scope next to it.

Measured independently

Measured or estimated by researchers or analysts with no stake in the result.

ecogpt.co estimate

Our own estimate from published factors. The working is shown next to the number; treat it as an order of magnitude, not a measurement.

Derived

Arithmetic on a disclosed or independent input using a published conversion factor.

Not disclosed

The company publishes nothing on this and no defensible estimate exists.

Pillars and weights

A company's score is the weighted sum of six pillar scores, each 0–100

Energy efficiency 20% How much electricity one answer costs.	Energy per prompt — disclosed, independently measured, or inferred from model architecture. 85–95Disclosed, measured per-prompt energy ≤ 0.25 Wh with methodology 65–75Disclosed figure without methodology, or independent estimate 0.25–0.5 Wh 60–75Open weights enabling independent measurement, no first-party figure 35–50Closed frontier models, nothing disclosed 30–40Image generation at several Wh per output, nothing disclosed
Carbon intensity 25% How clean that electricity is, and whether total emissions are rising.	Where the electricity comes from: hourly carbon-free share, grid mix, on-site fossil generation and the direction of absolute emissions. 80–90Low-carbon grid (< 100 gCO₂/kWh) or ≥ 80% hourly CFE 55–65100% annual matching plus a 24/7 goal; absolute emissions rising 30–45Grid-dependent in fossil-heavy regions, renewable certificates only 10–25On-site fossil generation for primary power
Water 15% How much water the data centres consume, and whether it is replaced.	Fleet water-use effectiveness, consumption trend and replenishment. 80–90WUE ≤ 0.2 L/kWh or zero-water cooling, replenishment on track 65–75WUE ≈ 0.3 L/kWh with a water-positive programme 50–60Rising consumption, partial replenishment, or WUE above 1 L/kWh 30–40Nothing disclosed
Transparency 20% How much the company actually publishes.	Eight disclosure checks, from per-prompt energy to third-party verification. 0–100Share of the nine-item disclosure checklist met (partial = half credit), plus up to 10 points for third-party review
Commitments 10% Whether it has dated, measurable climate and water targets.	Dated, measurable targets — and whether the company is moving toward or away from them. 80–90Dated net-zero, 24/7 CFE and water-positive targets, on track 55–70Net-zero 2040 or partial targets 10–25No public targets −10Deduction when the company itself reports moving away from a target
Conduct 10% How it behaves: permits, lawsuits, gas turbines, community water.	Permits, enforcement actions, community conflicts and clean-firm procurement choices. 75–90Clean-firm procurement, LCA leadership, no disputes 55–65No enforcement actions on record 30–45New fossil capacity secured for AI load; records withheld; permits that skip public comment ≤ 15Permit violations, pending Clean Air Act litigation

The nine disclosure checks

1.Per-prompt energy disclosed
2.Per-prompt water disclosed
3.Training emissions in model cards
4.Annual environmental report
5.Location-based Scope 2 reported
6.Data-center locations public
7.Water-use effectiveness reported
8.Third-party verified AI figures
9.Hardware reuse or e-waste reported

Confidence grades

Each pillar carries a grade; the company grade is the weight-averaged result

AHigh confidence

Company-disclosed with methodology, or independently measured, and consistent across sources.

BGood confidence

Disclosed at company level (annual report, model card) but not AI-specific, or disclosed without third-party review.

CMostly estimates

Derived from hosting partners, architecture or press reporting rather than the company's own numbers.

DLittle data

No disclosure of any kind; scored on what is publicly observable.

Estimates: how every blank is filled

166 of 210 key figures are estimates or derivations; 44 are disclosed or measured

Most companies in the index publish nothing per prompt and many publish nothing at all. Rather than leave a dash, every figure is filled with an estimate whose working is printed next to it (“How we got this number”) and labelled ecogpt.co estimate or Derived. Estimates are order-of-magnitude figures, typically within a factor of three, and they never raise a company’s score: scoring rewards disclosure, and an estimated figure earns the same confidence grade as no figure.

Energy per prompt	Anchored on the closest measured analogue: 0.3 Wh for a closed frontier model (Epoch's GPT-4o estimate), scaled by active parameters for mixture-of-experts models, 0.04 Wh for sub-10B models (Luccioni et al. median), 2–3× for retrieval-augmented search, and 1.6–3.6 Wh for a 1024-px image.
Water per prompt	Energy × the fleet's water-use effectiveness (WUE), on-site cooling only — the same boundary Google and OpenAI use. The full-scope figure, adding 1.8 L per kWh of electricity generated, is given in the note.
Carbon per prompt	Energy × the average carbon intensity of the grid where the prompts are served (location-based). Renewable certificates are not netted off; the note says so.
Annual emissions (non-reporters)	Fleet power in MW × 8,760 hours × 80% utilisation × grid intensity. Compute only; nothing else the company does is counted, so the figure is not comparable with a full corporate inventory.
Fleet power	Disclosed where available. Otherwise annual electricity ÷ 8,760 hours (average draw), or accelerator count × 1.3 kW all-in.
Accelerator count	Company statements and filings first, Epoch AI's H100-equivalent estimates second, and only then an order-of-magnitude estimate from announced capacity or capital spending — always with the uncertainty stated.
WUE and clean electricity (non-reporters)	Inherited from the host (AWS 0.15 L/kWh, Azure 0.30, Google 1.15) when a company runs on someone else's cloud; otherwise a stated mid-range assumption for the region and cooling type.

Grid carbon intensities used (location-based, gCO₂/kWh)

world average grid: 480
US grid: 390
EU grid: 240
French grid: 55
Chinese grid: 580
Texas grid plus on-site gas: 420
on-site gas turbines: 500
Nordic grid: 30

Hardware & materials

The factors behind the chips, copper and e-waste figures on every profile

Sustainability reports count electricity and water. The hardware behind AI has a footprint of its own — the carbon emitted making chips and servers, the copper in the building, the diesel generators out back, the e-waste at end of life — and almost nobody reports it. Each profile carries nine hardware figures. Where a company or analyst has published a number it is labelled as such; the rest are arithmetic on those inputs using the factors below, each of which cites its source.

Carbon baked into one AI accelerator and its share of the server 514 kg CO₂e per accelerator slot	164 kg CO₂e per H100 from NVIDIA's product carbon footprint, plus 350 kg for its share of the host server (1,400 kg per four-GPU node in the Jean Zay life-cycle assessment). arXiv 2607.05408 NVIDIA developer blog
Server mass per accelerator 17 kg of hardware per accelerator	A DGX H100 weighs 130 kg for eight GPUs (16.3 kg each); a GB200 NVL72 rack weighs about 1,360 kg for 72 GPUs (18.9 kg each). We use 17 kg. NVIDIA The Register
Power per accelerator including cooling and networking 1.3 kW per accelerator, all-in	xAI's Colossus runs roughly 230,000 accelerators on about 300 MW (SemiAnalysis), i.e. 1.3 kW each once hosts, networking and cooling are counted. SemiAnalysis
Copper per megawatt of data-centre capacity 27 tonnes of copper per MW	Microsoft's Chicago data centre used 2,177 t of copper, about 27 t per MW (BHP). S&P Global and Bloomberg Intelligence put AI-era sites at 20–40 t/MW. BHP S&P Global Market Intelligence
Backup generator size 2.6 MW per backup generator	Virginia has permitted about 10,500 data-centre generators totalling 27 GW (2.6 MW each); Stargate Abilene's permit lists 62 units for 170 MW (2.7 MW each). Latitude Media Latitude Media / Floodlight
Average draw as a share of installed power 0.8 share of nameplate	AI clusters run close to flat out; we assume 80% of nameplate capacity on average when converting megawatts to annual electricity. SemiAnalysis
Water evaporated generating electricity 1.8 L per kWh generated	US thermoelectric plants evaporate about 0.47 US gallons (1.8 L) per kWh delivered (NREL). Added to on-site cooling when a full-scope water figure is wanted. NREL via EPA

What each figure means

AI chips in serviceGPUs, TPUs or custom accelerators the company runs. Analyst figures are in H100-equivalents, company figures in physical chips.
Power drawMegawatts the AI fleet pulls. One megawatt runs roughly 750 US homes.
Electricity a yearTerawatt-hours per year. One TWh is a year of electricity for about 95,000 US homes.
Carbon baked into the hardwareCO₂e emitted making the chips and servers before they answer a single prompt.
Future e-wasteTonnes of server hardware that will be retired in three to five years.
Copper in the buildTonnes of copper in cabling, busbars and transformers.
Fossil generators on siteDiesel backup units or gas turbines at the data centres, permitted or observed.
Hardware reused or recycledShare of retired servers and parts given a second life.
Mineral depletionKilograms of antimony-equivalent, the life-cycle unit for scarce minerals consumed.

Scope matters

Why two honest numbers can differ 170×

Google’s 0.26 mL per prompt counts the water its own data centres consume for cooling. Mistral’s 45 mL per response also counts the water consumed generating the electricity, following a full life-cycle method. Neither is wrong; they answer different questions. The index records the boundary next to every figure, scores disclosure of the boundary itself, and never compares numbers across boundaries without saying so.

Emissions carry the same caveat: market-based totals net out renewable certificates and power-purchase agreements, location-based totals do not. Where both are reported (Meta, Microsoft, Google) the profile shows both. Estimated annual emissions for non-reporters cover the AI fleet only and are location-based.

Glossary

The jargon, in one line each

Scope 1: Emissions the company makes directly, e.g. gas turbines or generators it runs.
Scope 2: Emissions from the electricity it buys.
Scope 3: Everything else in the value chain: making chips and servers, building data centres, suppliers.
market-based: Scope 2 after subtracting renewable certificates and power contracts the company bought.
location-based: Scope 2 using the real average mix of the local grid, no certificates.
PUE: Power usage effectiveness: total facility power ÷ server power. 1.0 is perfect; 1.1 is excellent.
WUE: Water usage effectiveness: litres of water consumed per kWh of server electricity.
24/7 CFE: Carbon-free electricity matched every hour on the same grid, not just averaged over the year.
H100-equivalent: A way to compare different chips by converting them to the compute of one NVIDIA H100.
Sb-eq: Antimony-equivalent: the life-cycle assessment unit for depletion of scarce minerals.
MoE: Mixture of experts: a model that only activates a fraction of its parameters per token, which cuts energy per prompt.
LCA: Life-cycle assessment: an audited accounting of impacts from manufacturing through use to disposal.

What the index is not

Limits worth knowing before you quote a score

Not a measure of absolute impact. A small company with a perfect score can still have a footprint; a hyperscaler with a middling score dominates in absolute terms. Use the hardware figures, the calculator and analytics for scale.
Not real-time. Scores are recomputed monthly or when a material disclosure lands; each profile shows its last verification date.
Not independent measurement. Where companies disclose, we take their figures and label them disclosed; where third parties measure, we label that independent; where nobody has, we estimate and say so. We do not run benchmarks ourselves.
Not legal findings. Incident entries summarise public reporting and filings and link to them; allegations are described as allegations.

Independence and corrections

Who makes this and how to challenge it

The index is published by EcoGPT, a Delaware public benefit corporation that also runs an AI assistant at ecogpt.com. That is a conflict of interest, so EcoGPT is not ranked. It does not accept payment from any company it ranks. The newsletter is free.

Companies and readers can request a correction or supply new disclosures at team@ecogpt.com. A disclosed figure always replaces an estimate. Material corrections are noted on the affected profile with the date of change. Every company in the index is offered right of reply before a new edition is published.